IPR-GAN: Identity Preserving Representation GAN for Multi-view Face Synthesis and Recognition

Abstract

Photo-realistic and identity preserving multi-view synthesis from a single face image is a challenging and essential problem. The common challenges faced in multi-view face synthesis are that the serious appearance distortion suffering from face synthesis and the generated face images may keep “incomplete” identity information due to a single-pathway encoder-decoder network. This paper proposes Identity Preserving Representation Generative Adversarial Network (IPR-GAN) for photo-realistic multi-view face synthesis. IPR-GAN combats the challenging synthesis problems with a recognizing while generating framework and reserves the postural invariance identity data for downstream tasks like face recognition and pose estimation. Exhaustive experiments substantiate that the proposed method not only represents the improvement of multi-view face synthesis on visual realism, but also preserves identity information for face recognition.